Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation

TL;DR

This survey comprehensively reviews Attention Sink in Transformers, covering its utilization, interpretation, and mitigation strategies to improve model interpretability and performance.

Contribution

It is the first systematic survey that consolidates research on Attention Sink, clarifies key concepts, and guides future research directions.

Findings

Identifies key dimensions of Attention Sink research

Provides a structured overview of utilization, interpretation, and mitigation

Offers guidance for managing Attention Sink in Transformers

Abstract

As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work provides a pivotal contribution by clarifying key concepts and guiding researchers through the evolution and trends of the field. We envision this survey as a definitive resource, empowering researchers and practitioners to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.